Some quesitons: Training accuracy is too low

[guaguaxyf]

Hi, when I train and test in AB, I found some mistakes in my result. Can you tell me the reason?

[yangmin666]

I suggest you test on our public trained models of AB in 6fps to check whether there is something wrong with your data

[guaguaxyf]

Thanks for your reply! I also have some questions: 1.What is the purpose of this file and where is it?

2.Does the source code load the pretrained model, and if so, where is it? If not, what should I do? Looking forward to your help.

[yangmin666]

Thanks for your reply! I also have some questions: 1.What is the purpose of this file and where is it? 2.Does the source code load the pretrained model, and if so, where is it? If not, what should I do? Looking forward to your help.

1. It is only a output file in this framework which is not used, you can ignore it. 2. Of course this model load pretrained model, we use SlowOnly provided from facebookresearch. It seems that some users do not know how to use this pretrained model. We will supplement how to use it.

[guaguaxyf]

I have trained my data in thumos14_192_AB and thumos14_192_AF, but the value of mAP does not reach the level of paper，can you give me some suggestions? By the way, the checkpoint section in the readme gives the Baidu network disk link of the two files, what is it used for?

[yangmin666]

I have trained my data in thumos14_192_AB and thumos14_192_AF, but the value of mAP does not reach the level of paper，can you give me some suggestions? By the way, the checkpoint section in the readme gives the Baidu network disk link of the two files, what is it used for?

1: You have not used prerained model SlowOnly 2: Two checkpoint files are used to test THUMOS in 6 fps config.

[guaguaxyf]

I will appreciate if you supplement how to use the pretrained model SlowOnly early.

[yangmin666]

I will appreciate if you supplement how to use the pretrained model SlowOnly early. Pretrained model SlowOnly is updated.

[guaguaxyf]

Hi，where is ～/.cache/toch/hub？

[yangmin666]

Hi，where is ～/.cache/toch/hub？

If you use LINUX, you can find it in your home. Such as: /home/user/

[yangmin666]

Hi，where is ～/.cache/toch/hub？

[guaguaxyf]

Do I need to add code about the pretrained model SlowOnly to configs?

[yangmin666]

Do I need to add code about the pretrained model SlowOnly to configs?

No, you don't have to.

[guaguaxyf]

After I replaced the pretrained model as you suggested, the test results are still very low especially in AB 6ftps.

[yangmin666]

After I replaced the pretrained model as you suggested, the test results are still very low especially in AB 6ftps.

Inference operation has nothing to do with pretrained model. Do you have loaded the right .pth file and use the right config. I test this 6fps AB model in my machine and it performs well.

[guaguaxyf]

I think my configs are right, the right .pth file means?

[yangmin666]

After I replaced the pretrained model as you suggested, the test results are still very low especially in AB 6ftps.

I suggest you check your THUMOS frames data first. For example, each frame is 320x180, and 6fps is matched.

[yangmin666]

I think my configs are right, the right .pth file means?

epoch_300_weights.pth

[guaguaxyf]

I think my configs are right, the right .pth file means?

epoch_300_weights.pth

Using _CUDA_VISIBLE_DEVICES=0 python tools/thumos/test_ab.py configs/trainval/basictad/thumos14/basictad_slowonly_e700_thumos14_rgb_192win_anchor_based.py anchor_based_6fps/ epoch_300_weights.pth_ is fine to reference the test results of this weight file. But there is a problem with the test results of the 12 weight files produced by my training.

[yangmin666]

I think my configs are right, the right .pth file means?

epoch_300_weights.pth

Using _CUDA_VISIBLE_DEVICES=0 python tools/thumos/test_ab.py configs/trainval/basictad/thumos14/basictad_slowonly_e700_thumos14_rgb_192win_anchor_based.py anchor_based_6fps/ epoch_300_weights.pth_ is fine to reference the test results of this weight file. But there is a problem with the test results of the 12 weight files produced by my training.

"But there is a problem with the test results of the 12 weight files produced by my training." You mean that 300_weight.pth provides the right answer. But you can not train your own results as well as 300_weight.pth right?

[guaguaxyf]

I think my configs are right, the right .pth file means?

epoch_300_weights.pth

Using _CUDA_VISIBLE_DEVICES=0 python tools/thumos/test_ab.py configs/trainval/basictad/thumos14/basictad_slowonly_e700_thumos14_rgb_192win_anchor_based.py anchor_based_6fps/ epoch_300_weights.pth_ is fine to reference the test results of this weight file. But there is a problem with the test results of the 12 weight files produced by my training.

"But there is a problem with the test results of the 12 weight files produced by my training." You mean that 300_weight.pth provides the right answer. But you can not train your own results as well as 300_weight.pth right?

Yes, the pic is the test results in epoch_300_weights.pth. But when I tested in my 12 weight files, the answer were wrong.

[yangmin666]

AB 6fps has the result of 58.4. 54.6 is too low. It seems like you use 3fps data.

[shiyi-z]

hi! did you fix this issue? i have the same problem. the accuracy of the AF 6fps is 56.3. i used the checkpoint provided by baidu cloud in the readme and the data is 6fps.

[guaguaxyf]

hi! did you fix this issue? i have the same problem. the accuracy of the AF 6fps is 56.3. i used the checkpoint provided by baidu cloud in the readme and the data is 6fps.

no. My results and yours are basically the same, the accuracy under the same conditions is rounded to 56.3. only AF can run through. when testing AB, the test ends with task 53/212.